Convex Max-Product over Compact Sets for Protein Folding

Abstract

In this paper we present an approach to inference in graphical models with mixture of discrete and bounded continuous variables. In particular, we extend convex max-product to deal with these hybrid models and derive the conditions under which our approach is guaranteed to produce the MAP assignment. When dealing with continuous variables the messages are functions. We investigate a multi-grid approach which can be viewed as a piecewise constant representation of these messages. While this approach provides the theoretical guarantees it is not very practical. Inspired by this, we further propose a particle convex max-product algorithm that significantly outperforms existing particle methods in the task of protein folding and performs comparable to the state-of-the art while using a smaller amount of prior knowledge.